The Key Technique Research Of 3D OCT Fundus Image Processing System

As China's aging population, the increase in eye-prone industry, people's lack of awareness of eye protection and other reasons, the probability of the incidence of eye disease is increasing, and how to effectively detect early retinal diseases has become a serious issue which requires urgent solution. Along with the development of fundus photography, it's possible to diagnosis eye diseases by analyzing several retinal features.The fundus is the interior surface of the eye, which consist of retina, retinal vessels, optic nerve heads, optic nerve fibers, retinal macular and choroid. The lesions of these parts all are all called eye diseases. Fundus images could be obtained through various means. According to the access method, fundus images can be divided into IR(Infrared Radiation) images, ultrasound images and OCT(Optical Coherence Tomography) images. By detecting several fundus features, experienced doctors can diagnose the patient's eye diseases, such as glaucoma, diabetes, kidney disease, hypertension, bacterial diseases, parasite, and brain diseases. And the fundus features information mainly include the location and size of the optic disc, the inner retinal thickness, the location of fovea, vascular network, and the location of macular.The main content of this paper is to study three key technique of OCT fundus image processing, including retinal vascular network extraction, fovea positioning and layer segmentation.1) Firstly, this paper introduces an improved automatic method, which is based on vascular centerline extraction and morphological reconstruction algorithms, for retinal vascular network extraction. The algorithm starts with the extraction of vessel centerlines, which are used as guidelines for the subsequent vessel filling phase. The final segmentation is obtained using an iterative region growing method that integrates the contents of several binary images resulting from vessel width dependent morphological filters. On the basis of previous study, a multi-scale morphological details enhancement module is added to the original workflow in order to preserve some unobvious connected blood vessels. A comparison experiment finally proves that the improved algorithm can reach a higher segmentation accuracy.2) Secondly, the paper proposes a new light graph based algorithm for fovea location. For a retinal tomography image, we firstly calculate the light graph from the region which is 200 pixels upon the shortest path of the source image, then use an iterative polynomial fitting algorithm to obtain the fitting curve on the light graph, finally positioning the location of fovea by quantifying the fovea features. The algorithm can be extended to three-dimensional situation. Through experiments on a large amount of data (128×3 OCT images), we found that positioning accuracy of the algorithm can reach 95%. What we need to highlight is that the algorithm could be applied to the positioning of the optic disc, if the appropriate changes are made. Experiments on current data has confirmed that.3) In order to finish layer segmentation, our algorithm firstly divide the OCT fundus image into several vascular impact and non-vascular impact regions, then in order to keep boundary information and reduce noise, we apply bilateral filtering and median filtering to the non-vascular impact regions; Canny operator and gradient information are finally combined to finish the segmentation. As the vascular impacted regions of OCT fundus images with little layer information, which will affect the layer segmentation, this paper presents a iterative polynomial curve fitting algorithm to separate the vascular impacted regions and non-vascular impacted regions. The experiment result shows that the number of segmented layers is still too small, and the accuracy is also not high enough. Even though the algorithm need to be improved, the research has Laid the foundation for the follow-up study.